Comments on: Mellanox officially outs NVIDIA GPUDirect, NVIDIA says only the first step http://insidehpc.com/2010/05/28/mellanox-officially-outs-nvidia-gpudirect-nvidia-says-only-the-first-step/ HPC News Without the Noise for Supercomputing Professionals | insideHPC Sun, 09 Jun 2013 01:54:13 +0000 hourly 1 http://wordpress.org/?v=3.1.1 By: PeterW http://insidehpc.com/2010/05/28/mellanox-officially-outs-nvidia-gpudirect-nvidia-says-only-the-first-step/#comment-232565 PeterW Mon, 14 Jun 2010 13:24:57 +0000 http://insidehpc.com/?p=11630#comment-232565 Hello, Greg, you were saying "Peer-to-peer transfers over PCIe, like from SSD to a Tesla card, have been around for a while". May I ask you one could achieve Peer-to-peer transfers over PCIe from say a FGPA card and a Tesla? Cheers, peter Hello,

Greg, you were saying “Peer-to-peer transfers over PCIe, like from SSD to a Tesla card, have been around for a while”. May I ask you one could achieve Peer-to-peer transfers over PCIe from say a FGPA card and a Tesla?

Cheers,
peter

]]> By: John West http://insidehpc.com/2010/05/28/mellanox-officially-outs-nvidia-gpudirect-nvidia-says-only-the-first-step/#comment-229405 John West Sat, 29 May 2010 00:29:18 +0000 http://insidehpc.com/?p=11630#comment-229405 Greg - have you checked out the video briefing I linked to in the China post (http://www.mellanox.com/videos/GPU_Direct/GPU_Direct.html )? The approach for this case is outlined starting around 5:48, and going through about minute 8. As far as I can tell, there isn't a direct connection (no physical cable, for example) between the GPU and the HCA; it's just that both devices can now have direct access to the same pinned memory area so that the CPU itself doesn't have to be directly involved in the transfer (and no buffer copy is needed). Greg – have you checked out the video briefing I linked to in the China post (http://www.mellanox.com/videos/GPU_Direct/GPU_Direct.html )? The approach for this case is outlined starting around 5:48, and going through about minute 8. As far as I can tell, there isn’t a direct connection (no physical cable, for example) between the GPU and the HCA; it’s just that both devices can now have direct access to the same pinned memory area so that the CPU itself doesn’t have to be directly involved in the transfer (and no buffer copy is needed).

]]> By: Greg Pfister http://insidehpc.com/2010/05/28/mellanox-officially-outs-nvidia-gpudirect-nvidia-says-only-the-first-step/#comment-229331 Greg Pfister Fri, 28 May 2010 16:54:03 +0000 http://insidehpc.com/?p=11630#comment-229331 I'm a little confused by your SSD comment. Surely, since GPUDirect uses Mellanox' 40GB/s IB, this is a high-speed connection directly between Tesla cards, bypassing PCIe. Peer-to-peer transfers over PCIe, like from SSD to a Tesla card, have been around for a while. (Not necessarily used, but possible.) I’m a little confused by your SSD comment. Surely, since GPUDirect uses Mellanox’ 40GB/s IB, this is a high-speed connection directly between Tesla cards, bypassing PCIe. Peer-to-peer transfers over PCIe, like from SSD to a Tesla card, have been around for a while. (Not necessarily used, but possible.)

]]>